Matthew Berman reports an OpenAI announcement claiming that agents using an unreleased model produced a proof for the Navier-Stokes Millennium Prize problem. He introduces the equations through fluid-motion examples and explains why he considers the claimed result significant for AI research. The discussion does not establish independent validation of the proof or acceptance for the prize, so the announced solution remains a reported claim in this account.
Matthew Berman recounts an authorship dispute involving Tristan Buckmaster and Levent Alpöge, who had worked on a related fluid-mathematics problem with help from Codex. Matthew Berman attributes concerns about similar research directions, publication timing and possible access to private drafts to Tristan Buckmaster. He explicitly says that he does not know what happened and is relying on public posts and announcements; these concerns are allegations rather than a settled account.
Matthew Berman also presents OpenAI's response denying that its researchers or agents accessed the mathematicians' work before public release or accessed specific user data to solve the problem. He reports a separate qualification about whether deidentified product-usage data could have contributed to model improvement, together with Sébastien Bubeck's competing account of authorship and release coordination. The discussion preserves disagreement between the parties and does not establish that private drafts were used.
Matthew Berman connects the announcement to AI-assisted research and recursive self-improvement. He distinguishes tools that accelerate parts of a research workflow from a fully closed loop in which an AI proposes experiments, runs them, identifies improvements and updates itself repeatedly. His expectations about future mathematical discovery and scientific progress are extrapolations, not capabilities demonstrated or independently verified by the discussion.
Matthew Berman closes by raising questions about credit for discoveries and the risks of relying on model providers when doing private research or building a business. He argues that users should consider how research expertise and product-usage data interact with provider incentives. This is his interpretation of the dispute; the video does not establish a general rule that every hosted AI service trains on every user's private data or that future models will necessarily reproduce a customer's expertise.
Watch on YouTube
